The "scanned only once" is a bit misleading. A table scan is a horrendously expensive thing in SQL Server. At least up to SS2005, a table scan requires a read of all rows into a temporary table, then a read of the temporary table to find rows matching the join condition. So in the worst case, your query will read and write 1M rows, then try to match 2 rows to 1M rows, then delete the temporary table (that last bit is probably the cheapest part of the query). So if there are no usable indexes on B, you're just in a bad place.
In your second example, if B.Xyz is not indexed, the full table scan happens and there's a secondary match from 2 rows to 1000 rows - even less efficient. If B.Xyz is indexed, there should be an index lookup and a 2:1000 match - much faster & more efficient.
'course, this assumes the table stats are relatively current and no options are in effect that change how the optimizer works.
EDIT: is it possible for you to "unroll" the A rows and use them as a static condition in a no-JOIN query on B? We've used this in a couple of places in our application where we're joining small tables (<100 rows) to large (> 100M rows) ones to great effect.